A Medical Expense Data Processing System and Method Based on Multi-Source Data Fusion
By constructing a dual-time-series model based on ARIMA and LSTM and a multi-dimensional feature dataset, the problem of multi-dimensional dynamic management and control of medical insurance policies and clinical diagnosis and treatment was solved, realizing accurate traceability and closed-loop management of multi-level drug costs, and improving the accuracy and efficiency of medical insurance cost control.
Patent Information
- Application Number
- CN202611323297.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to balance multi-dimensional dynamic management of medical insurance policies and clinical diagnosis and treatment. Pre-warning thresholds are fixed and static, prediction models have poor accuracy, system functions are fragmented, data cannot be integrated across systems, there is a lack of multi-dimensional attribution and tracing capabilities, and feature extraction dimensions are limited, making it impossible to support multi-level and multi-scenario medical insurance cost control analysis.
By collecting data from HIS, electronic medical records, laboratory tests, and medical insurance, and combining CMI weights and medical insurance policies, an adaptive early warning threshold is constructed. ARIMA and LSTM dual time series models are used for data cleaning and feature extraction to construct a six-dimensional feature dataset, enabling multi-dimensional decomposition and traceability, and forming a closed-loop management of the entire process.
It enables accurate tracing of multi-dimensional cost anomalies, reduces false alarms and missed alarms, supports multi-level dynamic monitoring of drug costs, forms a closed-loop management system for the entire process, and meets the management needs of hospitals and departments.
Smart Images

Figure CN123120230A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical expense data processing technology, and in particular to a medical expense data processing system and method based on multi-source data fusion. Background Technology
[0002] The combined rigid regulation of rational drug use in clinical practice, coupled with medical insurance payment reform, necessitates hospitals to achieve refined management of drug costs throughout the entire process. However, existing methods struggle to simultaneously address the dynamic requirements of medical insurance policies and multi-dimensional clinical treatment management. Furthermore, hospital drug use, medical records, laboratory data, and medical insurance policies originate from multiple heterogeneous systems, resulting in inconsistent standards, numerous corrupted / missing values, and the inability to integrate and correlate data across systems. This hinders multi-dimensional joint analysis and the quantitative determination of illegal drug use. The core shortcomings of existing technologies are as follows: (1) The warning threshold is fixed and static, and is not adapted to CMI weights and medical insurance policies. This can easily lead to false warnings and missed warnings. It is impossible to distinguish between normal cost fluctuations and real anomalies based solely on absolute values.
[0003] (2) The prediction model is simple and has poor accuracy. It lacks the ability to trace the cause of multiple departments / drugs / diseases, and there are no standardized quantitative indicators to automatically identify unreasonable drug use such as overdosing and over-treatment. It relies on manual verification, which is inefficient and inconsistent in judgment.
[0004] (3) The system functions are fragmented. The data collection, analysis, early warning, intervention and reporting modules are isolated from each other. There is no hierarchical data distribution mechanism. The abnormal analysis results cannot be pushed to the corresponding intervention plan in a synchronous manner. The rectification effect has no feedback and cannot form a closed loop of control.
[0005] (4) The feature extraction dimensions are too thin, relying only on cost indicators and lacking a complete feature system such as drugs, departments, diseases, time series, and medication standards, making it difficult to adapt to multi-level and multi-scenario medical insurance cost control analysis.
[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.
[0008] According to one aspect of this application, a medical expense data processing method based on multi-source data fusion is provided, comprising: collecting drug expense-related data from the HIS system, including drug name, dosage, amount, and department affiliation information; setting dynamic early warning thresholds for cost, proportion, and growth rate, as well as quantitative judgment thresholds for irrational drug use, and determining the dispersion range between actual and predicted cost values, based on the cost control requirements of DRG / DIP medical insurance payment reform and clinical rational drug use evaluation standards; preprocessing the multi-source heterogeneous medical data, completing data integrity verification and format standardization processing, constructing a standardized basic medical expense dataset, and extracting feature dimensions such as cost, drug, department, disease, time series, and medication standardization for use in medical expense data processing. Cost Anomaly Identification and Medication Behavior Diagnosis: Combining the needs of anomaly identification, medication diagnosis, hierarchical control, and closed-loop management throughout the entire process, an absolute threshold early warning + relative value attribution analysis mode is adopted. An intelligent cost analysis strategy is constructed by integrating ARIMA and LSTM dual time-series prediction models. The source of anomaly contribution is traced through multi-dimensional decomposition, and the data output update cycle is matched with the visualization dashboard and automatic reporting. Data sets are split according to the dimensions of the whole hospital, departments, drug categories, and diseases, and multi-level control units are divided to form analysis data batches containing cost indicators, medication anomaly tags, compliance judgment results, and intervention plans. These data are simultaneously pushed to the early warning, intervention, and reporting application modules to support closed-loop management of early warning-diagnosis-intervention-feedback.
[0009] Another aspect of this application is a medical expense data processing system based on multi-source data fusion, used to execute executable instructions to perform the above-described medical expense data processing method based on multi-source data fusion.
[0010] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described medical expense data processing method based on multi-source data fusion by executing the executable instructions.
[0011] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for processing medical expense data based on multi-source data fusion.
[0012] The beneficial technical effects of this application are as follows: It collects four types of heterogeneous medical raw data: HIS, electronic medical records, laboratory data, and medical insurance data; then, combining CMI weights and medical insurance policies, it constructs an adaptive early warning threshold based on an ARIMA+LSTM dual-time-series model, along with quantitative judgment indicators such as dose deviation rate and treatment course exceeding limits; subsequently, it standardizes the data through cleaning, gap filling, and feature engineering, extracting six dimensions of features—cost, drug, department, disease, time series, and medication guidelines—to construct a unified dataset; it adopts a joint analysis mode of absolute threshold and relative attribution to decompose the contribution of abnormal costs from multiple dimensions to achieve accurate source tracing; finally, it packages data batches hierarchically by hospital, department, drug, and disease, and simultaneously pushes them to the early warning, intervention, and reporting modules, establishing a closed-loop management chain from early warning to diagnosis to intervention to feedback. Simultaneously, it provides multi-format visual dashboards to achieve hierarchical dynamic monitoring of drug costs, solving the technical pain points of traditional systems such as data fragmentation, rigid thresholds, difficulty in source tracing, and lack of closed-loop management.
[0013] This application aims to unify data standards across multiple systems, automatically clean and complete corrupted data and missing values, construct a complete six-dimensional feature dataset, eliminate heterogeneous data barriers, and support multi-dimensional horizontal comparative analysis. It employs a dual-time-series model combined with iterative adaptive thresholds for medical insurance and CMI parameters to distinguish between normal seasonal fluctuations and genuine anomalies, significantly reducing false alarms and missed alarms, while also quantitatively identifying various types of illegal drug use behaviors.
[0014] Furthermore, by breaking down abnormal contribution levels across three dimensions—department, drug, and disease—the core sources of drug cost increases can be quickly identified, significantly improving the targeted nature of control. A tiered distribution architecture automatically synchronizes indicators, abnormal tags, and intervention plans to various business modules. Data visualization links and displays real-time control dynamics, forming an iterative mechanism of analysis-rectification-feedback. Dashboards, specialty reports, and annual summaries provide multi-dimensional visual outputs, meeting the different monitoring needs of hospital and department-level management personnel and directly supporting medical insurance assessments and standardized clinical drug use management.
[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description
[0016] Figure 1 A flowchart of a medical expense data processing method based on multi-source data fusion provided in an embodiment of this application is shown; Figure 2 This is a system visualization dashboard that provides an annual summary display. Figure 3 This is a schematic diagram of a cardiology drug cost management analysis report; Figure 4 This is a schematic diagram of the drug cost structure analysis in a system visualization dashboard. Detailed Implementation
[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0018] In one implementation, Figure 1 A schematic flowchart of a medical expense data processing method based on multi-source data fusion according to an embodiment of this application is shown.
[0019] S101 collects drug cost-related data from the HIS system, including drug name, dosage, amount, and department affiliation information.
[0020] In one implementation, the HIS (Hospital Information System) medication data collection scope covers all outpatient and inpatient prescriptions, drug requisitions, drug settlements, and original business records of drug inventory turnover throughout the hospital. For example, it automatically captures the total monthly consumption of drugs such as piperacillin-tazobactam and amiodarone hydrochloride injection in the cardiology department, the number of prescriptions issued per transaction, drug settlement amounts, and year-on-year increases, and retains complete medication flow records for each patient, department, and drug throughout the entire time period, supporting subsequent statistics on total drug costs, drug ratio, and cost growth rate.
[0021] The electronic medical record data collection scope covers complete medical record text information, including patient disease records, treatment courses, clinical medication records, and physician medication diagnosis explanations. For example, it extracts treatment duration, medication course records, and clinical medication indication descriptions for patients in key departments such as oncology, respiratory medicine, and ICU, which are used to identify unreasonable behaviors such as medication without indication and medication beyond the prescribed course, and is matched with CMI disease group weighting calculation.
[0022] The scope of data collection for test results covers various laboratory test indicators, patient test times, and corresponding diagnostic and treatment items. For example, it simultaneously captures patients' blood routine and biochemical test results, matches them with concurrent prescription medication records, and uses this information to verify the matching degree between drug dosage and patient test indicators, and calculates the quantitative indicator of dosage deviation rate.
[0023] The scope of medical insurance policy data collection includes DRG / DIP disease group payment standards, medical insurance cost assessment targets, key monitored drug lists, annual medical insurance cost control constraints, and policy-identifying text data. For example, importing the hospital's annual drug cost control targets, medical insurance adjustment coefficients for various drugs, and lists of key high-priced controlled drugs provides policy-dimensional input parameters for ARIMA baseline prediction and LSTM residual correction, supporting iterative calculation of adaptive early warning thresholds.
[0024] S102 sets dynamic early warning thresholds for costs, proportions, and growth rates based on the cost control requirements of DRG / DIP medical insurance payment reform and the clinical rational drug use evaluation standards, as well as quantitative judgment thresholds for irrational drug use and the range of dispersion judgment between actual and predicted costs.
[0025] In one implementation, combining the mandatory cost control requirements of hospitals under the DRG / DIP medical insurance payment reform and the clinical rational drug use management and evaluation standards, based on the department's historical cost data, a baseline prediction range for costs, drug proportions, and cost growth rates is first generated using an ARIMA model. Then, an LSTM model is introduced to incorporate medical insurance policy identification data to complete residual correction. Combined with annual cost control targets, CMI disease group weights, and medical insurance adjustment coefficients, multi-dimensional parameters are iterated to output a multi-dimensional adaptively adjustable early warning threshold. First, a baseline calculation is completed based on the department's historical cost data. Then, the calculation results are corrected by combining the content of medical insurance policies, and multiple control parameters are superimposed to obtain dynamic early warning values suitable for each department. The input content includes long-term historical cost records of each department, the hospital's annual cost control targets, CMI weights for each disease, medical insurance drug adjustment standards, and medical insurance policy explanatory text.
[0026] First, the ARIMA model is used to calculate the basic ranges for costs, drug ratios, and cost growth rates. Then, these basic ranges, combined with medical insurance policy indicator data, are fed into an LSTM model to correct for deviations. Finally, the model iterates repeatedly, combining annual cost control targets, CMI disease group weights, and medical insurance adjustment coefficients to output early warning thresholds that can be adjusted according to business changes. Different departments treat different diseases and have varying levels of medical insurance control. By relying on the linkage of multiple parameters to adjust thresholds, a unified fixed value is no longer used for control judgments, thus adapting to the refined cost control needs of different departments.
[0027] For example, by reviewing the complete past medication cost records of the cardiology department, and combining the CMI weights of the corresponding cardiology diseases, the annual cost control indicators of the cardiology department, and the medical insurance adjustment coefficients of controlled drugs, the basic cost fluctuation range is first calculated. Then, the deviation of the range is corrected by combining the cardiovascular medical insurance control policies, and finally the cardiology-specific cost, drug ratio, and cost growth rate warning standards are obtained.
[0028] The adaptive early warning threshold employs an iterative algorithm combining weighted linear combinations and feedback correction. The first step calculates the initial threshold, which equals the cost control target weight multiplied by the corresponding cost control target threshold, plus the CMI weight multiplied by the threshold adjusted based on the CMI disease group weight, plus the policy weight multiplied by the threshold adjusted based on the medical insurance adjustment coefficient. The three weight coefficients are determined using the analytic hierarchy process (AHP). For example, in cardiology, the cost control target weight is 0.5, the CMI weight is 0.3, and the policy weight is 0.2. The second step introduces a feedback correction term. The next round's threshold equals the current round's threshold plus the learning rate multiplied by (the current round's error rate minus the target error rate). The error rate is the weighted sum of the false positive and false negative rates, with the target error rate set at 5% and the learning rate at 0.1. The iteration convergence condition is that the threshold change rate between two adjacent rounds is less than 1% or the maximum number of iterations (50) is reached. Threshold update triggers include changes in medical insurance policy, adjustments to the annual cost control target, and an early warning error rate exceeding 10% for three consecutive months.
[0029] The LSTM model employs a two-layer stacked network structure. The first LSTM hidden layer has a dimension of 64, and the second LSTM hidden layer has a dimension of 32. The time step is set to 12, which is equivalent to inputting historical data from the past 12 months. The activation function is ReLU, and the output layer is a fully connected layer. Model training uses the Adam optimizer with an initial learning rate of 0.001 and a mean squared error loss function. Training runs for 200 epochs, with an early stopping tolerance of 20 epochs. The training and validation sets are split in an 8:2 ratio. Dual-model fusion uses a serial residual correction mechanism. The specific formula is that the final predicted value equals the ARIMA baseline predicted value plus the LSTM residual predicted value, where the LSTM residual predicted value is the LSTM model's prediction of the ARIMA residual sequence (actual value minus ARIMA predicted value). Medical insurance policy identification data uses a hybrid encoding method. Policy types such as DRG payment reform, centralized procurement policies, and medical insurance catalog adjustments use one-hot encoding. Policy text descriptions are vectorized using word frequency inverse document frequency to reduce the dimension to 32, and then concatenated with the ARIMA residual sequence as the LSTM input features.
[0030] The ARIMA model uses the AIC criterion for order determination. By iterating through parameter combinations with autoregressive order p (0-5), differencing order d (0-2), and moving average order q (0-5), the order with the smallest AIC value is selected as the final model order. Taking the monthly cost series for cardiology as an example, the order determination result is ARIMA(2,1,1), meaning the autoregressive order p equals 2, the differencing order d equals 1, and the moving average order q equals 1. The ARIMA model output includes baseline predicted values and upper and lower bounds of the 95% confidence interval.
[0031] Based on a professional diagnostic knowledge base with built-in quantitative evaluation indicators such as CMI (Cost Elasticity Index), dosage deviation rate, and treatment duration exceeding limits, standardized judgment rules for unauthorized medication, overdose medication, and excessive treatment duration are constructed. Using the system's built-in professional diagnostic knowledge base, three types of quantitative evaluation indicators are extracted to establish unified judgment standards for these three types of clinical medication violations. Input content includes the patient's complete medical record, including drug dosage, duration of continuous administration, and relevant CMI values for the corresponding disease.
[0032] The system retrieves three quantitative indicators from the knowledge base: CMI (Cost Elasticity Index), dosage deviation rate, and treatment duration exceeding limits. Standardized judgment rules are then established for these three categories, corresponding to medication use without indication, overdose, and over-treatment duration, respectively. All judgment indicators are formulated in accordance with clinical rational drug use guidelines and medical insurance violation penalty requirements. This system can automatically identify non-compliant clinical medication behaviors, replacing manual verification of medical records.
[0033] For example, by retrieving complete diagnosis and treatment information of ICU cancer patients, calculating the cost elasticity coefficient by combining the corresponding disease CMI value, comparing the deviation range with the standard dosage, and statistically analyzing the overall dosing duration to determine whether it exceeds the standard treatment course, any indicator reaching the standard can identify the corresponding irrational drug use type.
[0034] For drugs with a narrow therapeutic window, such as amiodarone and warfarin, the patient's liver and kidney function test results (creatinine clearance rate, transaminase) are used as correction factors. The standard dose is adjusted individually before calculating the deviation rate. When the creatinine clearance rate is less than 30 ml / min, the standard dose is reduced by 50%. When the transaminase exceeds three times the upper limit of normal, the standard dose is reduced by 30%. The treatment duration exceeding the limit is determined based on the maximum treatment duration specified in the drug instructions, the treatment duration recommended by clinical guidelines, and historical medication data. The smaller value between the upper limit of the treatment duration specified in the drug instructions and the upper limit of the treatment duration recommended by clinical guidelines is taken as the baseline treatment duration. Then, the 95th percentile of the actual treatment duration distribution of the same disease and the same drug in the past 12 months is taken as the statistical treatment duration. The final treatment duration exceeding the limit is the larger value between the baseline treatment duration and the statistical treatment duration. When the number of consecutive days of medication administration (intervals not exceeding 48 hours are considered continuous treatment) exceeds the treatment duration exceeding the limit, it is determined as excessive medication use. Taking piperacillin-tazobactam for injection for community-acquired pneumonia as an example, the instructions for use state that the maximum course of treatment is 14 days, the guidelines recommend a course of treatment of 7 to 10 days, and the 95th percentile of historical data is 12 days. Therefore, the maximum course of treatment is 14 days. If a patient takes the medication continuously for more than 14 days, it is considered to be an over-course treatment.
[0035] The dose deviation rate quantifies the degree of deviation between the actual administered dose and the standard administered dose. The specific calculation formula is: dose deviation rate = (actual single-dose minus standard single-dose) divided by the standard single-dose multiplied by 100%, where both the actual and standard doses are normalized to body weight and the unit is milligrams per kilogram. The standard administered dose is derived from the system's built-in clinical drug knowledge base, integrating drug instructions, clinical practice guidelines, and expert consensus, indexed by three dimensions: generic name, indication, and route of administration. The dose conversion rules for different routes of administration are as follows: oral doses and intravenous bolus doses are converted based on drug bioavailability, while intravenous infusion doses are converted to a single equivalent dose based on the administration rate per unit time. The dose deviation rate judgment threshold is divided into three levels: less than or equal to 20% is considered reasonable; greater than 20% but less than or equal to 50% is considered mild overdose; and greater than 50% is considered severe overdose.
[0036] The CMI-cost elasticity coefficient is defined as the ratio of the cost change rate to the CMI change rate, reflecting the degree to which changes in disease structure affect costs. The specific calculation formula is: elasticity coefficient equals (cost change divided by base period cost) divided by (CMI change divided by base period CMI). This coefficient is determined using a disease-specific linear regression method. Using monthly cost and CMI data from the past 24 months as samples, a regression equation is fitted for each disease, where the logarithm of cost equals the intercept plus the elasticity coefficient multiplied by the logarithm of CMI. The regression coefficient is the CMI-cost elasticity coefficient for that disease. The coefficient typically ranges from 0.5 to 3.0. An elasticity coefficient greater than 1.5 indicates high sensitivity to changes in disease structure, while a value less than 0.8 indicates insensitivity. The logic for determining no-indication medication use is that if a patient's medication record shows a CMI-cost elasticity coefficient greater than 2.0 and no corresponding clinical indication description for the drug is found in the electronic medical record, the patient is considered to have suspected no-indication medication use.
[0037] This study integrates ARIMA and LSTM dual time-series prediction models to perform time-series cost fitting operations, calculates the dispersion between actual cost values and model predictions, and sets a critical boundary for dispersion as the basis for identifying cost anomalies. The two time-series models, ARIMA and LSTM, are used to simultaneously fit periodic cost data, and the difference between predicted and actual costs is compared to define critical criteria for identifying cost anomalies.
[0038] The input consists of real drug cost records for the entire hospital and each department at different time periods. A dual-time-series model is used to fit the periodic costs, and the dispersion of the difference between the actual costs and the model-predicted costs for the same period is calculated. A critical value for the dispersion is set in accordance with the hospital's medical insurance control requirements, serving as the boundary for identifying cost anomalies. Normal cost fluctuations caused by seasonal fluctuations and changes in disease structure are eliminated, and only data with fluctuations exceeding the critical standard are marked as anomalies, reducing meaningless warnings.
[0039] For example, by summarizing the actual drug cost data of the respiratory medicine department for the past five months, generating the predicted cost for the same period through a dual time series model, calculating the fluctuation difference between the actual and predicted costs each month, and determining that there is an abnormality in the control of drug costs for that month when the difference exceeds a preset threshold value.
[0040] The joint judgment logic of absolute threshold and dispersion adopts an AND relationship. The actual cost is judged as abnormal only when it simultaneously exceeds the absolute warning threshold and the absolute value of Z is greater than 2. This mechanism can effectively filter out absolute threshold exceedances caused by seasonal fluctuations and normal changes in disease structure, reducing the false alarm rate of simple absolute threshold warning by about 40%.
[0041] Cost dispersion is quantified using Z-score standardized residuals. The specific formula is Z = (actual cost minus dual-model fusion prediction) divided by the historical residual standard deviation, which is calculated based on the predicted residual sequence over the past 24 months. The critical boundary for dispersion is set using a statistical distribution method. Based on the assumption of a normal distribution of historical residuals, Z = ±2 is taken as the boundary of the 95% confidence interval; that is, a Z absolute value greater than 2 is considered an abnormal cost. For departments with large cost fluctuations, such as the ICU and oncology departments, department-specific differentiated thresholds are used. Based on the actual distribution of historical residuals within the department, the 97.5th percentile is used as the upper bound, and the 2.5th percentile as the lower bound, to avoid false alarms caused by a uniform threshold.
[0042] It features a combined absolute threshold early warning and relative value attribution analysis identification mode, integrating multi-dimensional analysis of dispersion anomalies by department, drug category, and disease type to support accurate tracing of abnormal drug costs. By combining threshold early warning and data attribution analysis, it breaks down the proportion of anomalies at three levels—department, drug category, and disease type—to pinpoint the core sources of cost anomalies. The input consists of abnormal cost data that has already been labeled with dispersion, accompanied by three business tags: department, drug category, and disease type.
[0043] Based on the overall discrete anomaly values of costs, the system breaks down the anomaly contribution percentage by department, drug category, and disease, quantifying the impact of different business dimensions on overall drug cost anomalies. Matching the hospital's tiered drug control model, managers can pinpoint cost increases caused by the entire hospital, a single department, a specific drug category, or a specific disease, accurately identifying key control areas.
[0044] For example, after determining that the hospital's monthly drug expenses were abnormal, the abnormal contribution percentages for oncology, proton pump inhibitor drugs, and gastrointestinal diseases were calculated separately. This confirmed that the increase in expenses was mainly due to the increased use of proton pump inhibitors in the oncology department, thus completing the precise tracing of the problem.
[0045] Anomaly contribution is calculated using a quantitative method based on cost increments. The specific formula is: contribution for a given dimension equals the cost increment for that dimension divided by the overall anomaly cost increment multiplied by 100%, where the cost increment is the actual cost minus the dual-model prediction. When multiple dimensions intersect, a marginal contribution uniform distribution strategy is adopted. For each cost record, its incremental percentage within the aggregation of each dimension is calculated separately for department, drug, and disease. The average of these three percentages is taken as the record's overall contribution, which is then aggregated across dimensions to avoid double counting caused by single-dimensional attribution. Dimensions with a contribution exceeding 30% are marked as primary contribution sources, and those exceeding 15% are marked as secondary contribution sources. Managers prioritize handling high-percentage anomalies based on their contribution ranking.
[0046] S103 performs preprocessing on multi-source heterogeneous medical data, completes data integrity verification and format standardization, constructs a standardized basic dataset of medical expenses, and extracts feature dimensions such as cost, drugs, departments, diseases, time series, and medication guidelines for use in cost anomaly identification and medication behavior diagnosis.
[0047] In one implementation, considering the inherent data defects of multi-source heterogeneous medical data such as HIS medication data, electronic medical records, test results, and medical insurance policies—including data disorder, missing indicators, and inconsistent business fields—a standardized preprocessing operation is uniformly performed on multiple types of original medical business data. This operation includes data cleaning, missing value completion, and feature engineering to remove invalid and dirty data, fill in gaps in business indicators, and unify the field formats across multiple systems. The four types of original medical data—HIS medication data, electronic medical records, test results, and medical insurance policies—generally suffer from disorganized records, missing indicators, and inconsistent system field standards. This approach provides clean and reliable foundational data for subsequent indicator calculations and medication compliance determination. The unified data cleaning, missing value completion, and feature engineering operation simultaneously removes duplicate, invalid, and dirty data, fills in missing business indicators, and unifies the field representation formats across multiple business systems.
[0048] Data collection nodes are deployed in the hospital’s outpatient, inpatient, laboratory, and medical insurance business systems. The raw data from each system is synchronously aggregated to the preprocessing entry point. Unified cleaning rules are set for duplicate prescription records, blank treatment courses in medical records, and differences in drug name fields. Records with missing dosing duration, disease CMI value, and drug settlement amount are batch-filled with the corresponding business benchmark values.
[0049] Medication records, medical records, laboratory reports, and medical insurance control documents are all included in the scope of standardization. Outpatient prescription data is centrally standardized by monthly batches, complete inpatient medical records are separately standardized by patient's single hospitalization cycle, and medical insurance policy documents are uniformly standardized and converted according to the annual control cycle.
[0050] The original standardized data is automatically bound to the business source identifier, which adopts a two-level coding rule. The first-level code distinguishes four types of data sources: HIS, electronic medical records, laboratory, and medical insurance. The second-level code distinguishes outpatient, inpatient, and annual policy sub-types. For example, outpatient HIS medication data is coded as HIS-MZ. The standardized data is automatically labeled with treatment time period tags to distinguish between daytime outpatient treatment and nighttime inpatient monitoring, thus completely retaining the original information of the entire medical business cycle.
[0051] Drug cost data is retained to two decimal places, treatment duration is recorded in calendar days, CMI weight ranges from 0.5 to 3.0, dose deviation rate ranges from 0% to 100%, and missing indicators are marked with the source of completion. When duplicate or invalid data is identified, dirty data removal records are generated simultaneously.
[0052] This paper establishes standardized dataset construction specifications by combining DRG / DIP medical insurance cost control analysis requirements, drug cost anomaly identification objectives, and irrational drug use diagnosis scenarios. It also includes cross-system data alignment rules and unified indicator conversion logic. Based on the three business analysis objectives of DRG / DIP medical insurance cost control statistics, drug cost anomaly identification, and irrational clinical drug use diagnosis, a unified dataset construction standard is established to provide a unified statistical benchmark for cross-departmental and cross-disease horizontal cost comparisons. This is complemented by cross-system data alignment rules and unified indicator conversion logic across all business operations.
[0053] The hospital-wide, specialty, and single-disease-based three-tiered analysis units share the same dataset construction standards. Unified conversion standards are set for outpatient and inpatient drug cost division, high-priced monitored drug conversion, and CMI weight matching to eliminate data deviations caused by different departments' custom statistical rules.
[0054] Specialized supplementary rules for key management departments such as oncology, respiratory medicine, and ICU are configured separately. All departments adopt the same calculation logic for drug ratio, cost growth rate, and treatment course exceeding limits. The subdivision rules for specialties are only slightly adjusted for disease weights, without changing the calculation method of core indicators.
[0055] The integrated dataset is bound to two-level partition codes: department and disease. The first-level code distinguishes between the whole hospital and major specialty categories, while the second-level code distinguishes between specific diseases. For example, the code for gastrointestinal diseases in the oncology department is ON-GX. The dataset is automatically bound to the annual medical insurance management cycle label to distinguish between different DRG payment annual statistical batches, ensuring consistent data standards across years.
[0056] The drug ratio and cost growth rate indicators are uniformly retained to two decimal places. The CMI disease group weight values are synchronously embedded in each diagnosis and treatment data. When there are differences in the scope of cross-system alignment, a scope conversion log is automatically generated and kept for future reference.
[0057] A standardized medical expense dataset construction strategy is implemented based on a multi-dimensional business feature extraction mechanism, extracting six major feature dimensions: cost, drugs, departments, diseases, time series, and medication guidelines. This standardized business feature extraction mechanism completes layered data extraction and integration to form a standardized basic medical expense dataset. It fully extracts the six major business feature dimensions—cost, drugs, departments, diseases, time series, and medication guidelines—to support subsequent anomaly identification and comprehensive diagnosis of clinical medication behavior.
[0058] Feature extraction is performed in a fixed hierarchical order, prioritizing the extraction of cost features such as total monthly expenses and single drug consumption, followed by the extraction of five supporting features in sequence: drug category, affiliated department, corresponding disease, time period sequence, and treatment course and dosage. Each of the six features is bound to the corresponding original medical records.
[0059] For key monitored drugs such as proton pump inhibitors and blood products, we have separately enhanced feature markers and simultaneously extracted supporting information such as indications for use, duration of administration, and compliance of single dose, which are separately included in the feature dimension of drug use guidelines for rapid identification of irrational drug use.
[0060] The extracted feature dataset is encapsulated in layers according to management units. The first-level unit distinguishes between the entire hospital and departments, while the second-level unit distinguishes between drugs and diseases. For example, the batch code of key respiratory drugs is RE-DR. All features are automatically bound to monthly time-series labels, and the time-series segments are divided according to natural months to completely retain long-term cost fluctuation information.
[0061] Cost-related features are uniformly retained to two decimal places, while treatment course and dosage features are retained as integers. Each feature is synchronously associated with the original medical record and prescription traceability number. Missing feature items are automatically marked when feature extraction is missing, ensuring that subsequent model and judgment index calculations are complete and usable.
[0062] All completed data is simultaneously labeled with its source. The label fields include four categories: department median, standard CMI, average settlement price, and KNN imputation. Confidence levels are set to high, medium, and low. When participating in indicator statistics, the completed data has a weight of 0.8, while the original data has a weight of 1.0. The completed data is not used for model training; it is only used for statistical display of the complete dataset, ensuring the authenticity and reliability of the model training data.
[0063] Missing value completion employs a categorized, differentiated strategy, with the following specific rules: First, for missing medication duration, the median medication duration of patients with the same disease and medication in the same department is used. If the sample size of patients with the same disease and medication in the same department is less than 10, the median of the same disease and medication across the entire hospital is used. Second, for missing disease-specific CMI values, the standard CMI value of the DRG group corresponding to the patient's primary diagnosis is used. The standard CMI value is derived from the DRG grouping scheme and weight table published by the National Healthcare Security Administration. If the primary diagnosis cannot be matched to a DRG group, the average CMI value of the same department over the past 6 months is used. Third, for missing drug settlement amounts, the recent average settlement price of the same drug, specification, and batch is used. If there are no settlement records for the same batch within the past 30 days, the medical insurance price limit for the drug is used as the filling value. Fourth, for missing test results, the K-nearest neighbor method is used, selecting the average test results of 5 patients with the same disease in the same department, whose age difference is no more than 5 years, and who are of the same gender, for filling.
[0064] S104 combines the needs of anomaly identification, medication diagnosis, hierarchical control, and full-process closed-loop management. It adopts an absolute threshold early warning + relative value attribution analysis mode, integrates ARIMA and LSTM dual time series prediction models to build a cost intelligent analysis strategy, relies on multi-dimensional decomposition to achieve anomaly contribution tracing, and matches the data output update cycle of the visual dashboard and automatic report.
[0065] In one implementation, combining the business needs of anomaly identification of drug costs, intelligent diagnosis of irrational drug use, hierarchical management of the whole hospital and departments, and closed-loop management of early warning, diagnosis, intervention and feedback, the system relies on the joint identification mode of absolute threshold early warning and relative value attribution analysis to call the preprocessed standardized medical cost dataset, match the three-dimensional analysis decomposition boundary of department, drug category and disease, locate the cost statistics node corresponding to multi-source medical data, and divide the calculation range of cost anomaly contribution under each dimension.
[0066] The input is a standardized, multi-dimensional medical expense dataset. Each data point is pre-labeled with department, drug, and disease classification tags. The system defines independent calculation intervals for each of these three dimensions to break down the sources of overall drug cost anomalies. This distinguishes between overall cost fluctuations and anomalies in specific dimensions, avoiding the inability to pinpoint specific control targets by only identifying hospital-wide anomalies, thus adapting to the hospital's tiered drug cost management system. For example, when overall monthly drug costs are abnormal across the hospital, the system separately divides the calculation into three categories: oncology, proton pump inhibitors, and gastrointestinal diseases, calculating the percentage increase in costs for each category.
[0067] The engineering decomposition of abnormal contribution adopts a hierarchical calculation method based on cost increment. Specifically, the departmental contribution equals the cost increment of a certain department divided by the total hospital-wide cost increment multiplied by 100%, where the cost increment is the actual value minus the predicted value. Further decomposition at the departmental level involves drug and disease dimensions. Drug contribution equals the cost increment of a certain drug in that department divided by the total cost increment of that department multiplied by 100%, and disease contribution equals the cost increment of a certain disease in that department divided by the total cost increment of that department multiplied by 100%. Multi-dimensional cross-attribution employs a marginal contribution uniform allocation strategy. When a cost increment is simultaneously assigned to multiple dimension labels, it is allocated according to the marginal contribution ratio of each dimension to avoid double counting. The hierarchical decomposition order is: hospital-wide abnormality to departmental contribution, then to the designated key department, then to the drug contribution of that department, then to the designated key drug, then to the disease contribution of that drug, and finally to the designated key disease, forming a three-level traceability chain. Dimensions with a contribution exceeding 30% are marked as primary contribution sources, and those exceeding 15% are marked as secondary contribution sources.
[0068] The system categorizes and labels drug cost indicators, clinical medication behavior, medical insurance compliance information, and time-series consumption data across various dimensions. It integrates ARIMA and LSTM dual-time-series prediction models to calculate the dispersion between actual and predicted costs. Based on quantitative indicators such as CMI (cost elasticity coefficient), dose deviation rate, and treatment duration exceeding limits, it identifies irrational drug use. The system is adapted to a visual dashboard and intelligent analysis reports with a fixed update cycle, ensuring complete data consistency for indicators highlighted in red, multi-level data drill-down, and automatic report pushes. Furthermore, it unifies the classification and labeling of drug cost indicators, clinical medication behavior, medical insurance compliance information, and time-series consumption data by department, drug category, and disease. It uses ARIMA and LSTM dual-time-series models to perform cost fitting calculations, assessing the dispersion between actual and predicted costs. It uses CMI (cost elasticity coefficient), dose deviation rate, and treatment duration exceeding limits to identify irrational drug use. The output content is adapted to a dashboard and intelligent report with a fixed update cycle, ensuring data consistency for indicators highlighted in red, multi-level drill-down, and automatic report pushes.
[0069] The input is a time-series cost dataset with complete dimension annotation. The model consists of two layers: an ARIMA baseline fitting layer and an LSTM policy residual correction layer. These two layers are operated sequentially, and discrete numerical values combined with knowledge-based quantitative indicators simultaneously complete medication compliance screening. The time-series model eliminates normal cost fluctuations caused by seasons and diseases, while quantitative indicators replace manual verification of medical records. A fixed update schedule ensures that managers receive timely control information. For example, by summarizing five months of complete respiratory medicine treatment data, the two models generate standard drug cost ranges for the same period, calculate the actual cost dispersion, and simultaneously verify medication dosage and treatment duration to determine violations. All results are updated monthly in a specialist analysis report.
[0070] A multi-dimensional anomaly tracing and visualization output linkage mechanism is established. After the dataset completes the decomposition and calculation of the anomaly contribution of departments, drug categories, and diseases, hierarchical control analysis results are generated and arranged chronologically according to a preset monitoring cycle. These results are then pushed to the front-end visualization application module to display the real-time dynamics of drug cost control across the entire hospital, specific departments, and individual drug products.
[0071] The input carrier is the stratified analysis results of the anomaly rate calculation, with data compiled according to monthly and quarterly monitoring cycles. The data is pushed through three front-end carriers: hospital-wide dashboards, departmental dashboards, and automated analysis reports. This enables automatic linkage between anomaly calculation results and visualization terminals, eliminating the need for manual data export and organization. The stratified display meets the viewing needs of both hospital-level and departmental management personnel. For example, after completing the hospital-wide multi-dimensional traceability calculation of drug costs each quarter, the system automatically breaks down the analysis content for each department and simultaneously pushes the corresponding visualization interface, allowing direct viewing of the year-on-year cost increases for key drugs and changes in the total amount of irrational drug use.
[0072] S105 splits the dataset by hospital, department, drug category, and disease, divides it into multi-level control units, and forms batches of analytical data that include cost indicators, medication anomaly labels, compliance judgment results, and intervention plans. These data are then pushed to the early warning, intervention, and reporting application modules to support closed-loop management of early warning, diagnosis, intervention, and feedback.
[0073] In one implementation, based on DRG / DIP medical insurance cost control, tiered clinical management, and multi-module data carrying capacity, a tiered data distribution system supporting the entire closed-loop management of drug costs is taken as the top-level goal of dataset splitting. The overall design is based on DRG / DIP medical insurance payment control requirements, in-hospital tiered clinical management needs, and the data carrying capacity limit of the front-end early warning and intervention report module. The top-level design goal for dataset splitting is to establish a tiered data distribution system supporting complete closed-loop drug cost management. All data splitting, encapsulation, and push logic is built around the implementation of the entire drug cost management business.
[0074] The input base is a standardized analysis dataset after multidimensional anomaly tracing. The overall architecture design is tailored to the hierarchical viewing needs of hospital and department-level management personnel, while also taking into account the monthly and quarterly periodic statistical reporting requirements of medical insurance. It uniformly coordinates the data receiving capacity of each application module. This avoids chaotic data splitting standards and pushing data volumes that exceed the capacity of the front-end modules. The unified top-level design ensures that the entire data flow adapts to the closed-loop management process.
[0075] For example, hospitals need to simultaneously push drug cost analysis content to the overall hospital module, the cardiology module, and the monthly medical insurance report module. The top-level distribution system pre-matches the data receiving capabilities of the three types of modules and uniformly defines the overall rules for splitting the dataset, so that there will be no problems such as data overload or missing data in a single module.
[0076] The system divides data into four primary control units: hospital-wide, departmental, drug category, and disease type. These units further derive secondary data packaging units for cost collection, anomaly labeling, compliance accounting, and intervention plan matching. The first-level control units are divided according to four business dimensions: hospital-wide, departmental, drug category, and disease type. Each primary unit then derives four secondary data packaging units for cost collection, anomaly labeling, medication compliance accounting, and clinical intervention plan matching, achieving multi-dimensional, layered, and refined data packaging.
[0077] The input data consists of complete analytical data for anomaly cost breakdown and medication compliance determination. First, it's segmented into four major dimensions. Then, within each dimension, the data is further broken down into four sub-business items and encapsulated separately, forming structured data batches. The first-level units meet the needs of macro-level hierarchical management and viewing, while the second-level encapsulation units separately break down different business information. Subsequent front-end modules can retrieve single-type data as needed, eliminating the need to load all redundant content at once.
[0078] For example, taking the primary control unit of the oncology department as an example, it further encapsulates four secondary data units: monthly total cost data of the oncology department, abnormal use labels of anti-tumor drugs, compliance judgment results of various chemotherapy drugs, and high-priced drug alternative intervention programs, forming a complete analysis data package exclusive to the oncology department.
[0079] A progressive, hierarchical data distribution architecture is established, consisting of a top-level objective, a primary control unit, and a secondary encapsulation unit. This architecture clarifies data input at each level, hierarchical relationships, multi-module push mechanisms, and timing constraints. Complete batches of encapsulated drug cost analysis data are synchronously pushed to the early warning, intervention, and reporting modules. This achieves closed-loop management of early warning, diagnosis, intervention, and feedback through multiple clinical intervention methods. A three-tiered, progressive data distribution architecture is also established, defining the data sources, hierarchical relationships, data push channels for multiple application modules, and timing constraints for each layer. The integrated complete set of drug cost analysis data is synchronously pushed to the three front-end business modules: early warning, intervention, and reporting. This achieves closed-loop management of early warning, diagnosis, intervention, and feedback through a complete set of clinical intervention measures, including pre-prescription interception, medication reminders during use, post-treatment review, and high-priced drug substitution.
[0080] The input consists of standardized analysis data batches that have been layered and encapsulated. The architecture manages the entire data chain from top to bottom, including data splitting, packaging, and scheduled push notifications. The push timing matches the monthly and quarterly medical insurance monitoring cycles, and the three types of modules can simultaneously receive data from their respective dimensions. This progressive architecture streamlines the data flow sequence and push timing rules, establishing a complete pathway from data analysis results to clinical intervention. This enables the direct push of corresponding intervention plans for implementation and rectification upon the discovery of cost anomalies.
[0081] For example, after the hospital-wide data is packaged and layered each quarter, it is pushed synchronously to the early warning module, the clinical intervention operation module, and the quarterly drug expense report module according to the preset time sequence. The reports also carry abnormal drug use in each department and corresponding alternative intervention plans. After viewing the reports, managers can directly rely on the intervention methods to standardize clinical drug use.
[0082] The personalized adjustment of the intervention plan is based on the department level and the doctor's professional title, setting an intervention intensity coefficient. The coefficient is 1.2 for key controlled departments (oncology, ICU, respiratory department), 1.0 for general departments, 0.8 for senior doctors, 1.0 for intermediate doctors, and 1.2 for junior doctors. The coefficient, combined with the basic intervention intensity, determines the final implementation strength of the intervention measures. The intervention effect feedback evaluation adopts the before-and-after comparison method. For each intervention event, the change rate of the corresponding abnormal indicators within 30 days after the intervention is calculated. The change rate is equal to (pre-intervention indicator value minus post-intervention indicator value) divided by the pre-intervention indicator value multiplied by 100%. A change rate greater than or equal to 20% is evaluated as effective, 0 to 20% is evaluated as partially effective, and less than 0% is evaluated as ineffective. The evaluation results are fed back into the intervention plan database to optimize rule matching weights and intervention measure combinations.
[0083] The matching rules between abnormality types and intervention plans are based on a rule engine mapping table. The specific mapping relationships are as follows: Overdosing (dosage deviation rate greater than 50%) corresponds to pre-prescription interception plus pharmacist in-process telephone reminders plus post-treatment special reviews, with a strong intervention intensity; Over-treatment (exceeding the treatment course limit) corresponds to pop-up reminders in the application of the pharmacy system plus department head meetings plus post-treatment reviews, with a medium intervention intensity; Medication without indication (CMI-cost elasticity coefficient greater than 2.0 and no indication in the medical record) corresponds to pre-prescription review interception plus pre-treatment communication with pharmacists plus monthly reports on irrational medication use, with a strong intervention intensity; Abnormal costs (absolute Z-value of dispersion greater than 2 and exceeding the absolute threshold) correspond to departmental drug cost warning letters plus dynamic adjustment of cost control targets plus quarterly pharmacy meeting reports, with a medium intervention intensity; Abnormal growth in the usage of key monitored drugs corresponds to recommendations for high-priced drug alternatives plus special reviews of drug use plus limited procurement, with a medium intervention intensity.
[0084] Intervention matching employs a rule-engine-based mapping and matching mechanism. The intervention database is constructed using a hybrid approach combining manual pre-configuration and historical case accumulation. The initial database consists of 12 categories of intervention rules pre-configured by clinical pharmacy experts based on documents such as the "Hospital Prescription Review Management Standards" and the "Guidelines for Clinical Application of Antimicrobial Drugs." Each rule category includes three elements: triggering conditions, intervention measures, and intervention intensity. During system operation, historical cases where the intervention effect is assessed as effective (corresponding abnormal indicators decrease by ≥20% after intervention) are automatically accumulated as candidate rules and incorporated into the official rule database after expert review.
[0085] like Figure 2 As shown, Figure 2This is a system visualization dashboard displaying an annual summary chart. The statistical period is from January 1, 2025 to December 31, 2025, and the statistical scope is a year-round summary. The upper right corner of the page features preview and report download buttons, supporting online viewing and local export of the report.
[0086] The first section displays the hospital's overall annual drug cost accounting results, fully detailing the total annual drug cost, year-on-year growth rate, and comparing it with the annual control growth rate target. It differentiates between outpatient and inpatient drug costs and their respective percentages, indicating compliance with targets. Simultaneously, it lists the actual values, corresponding target thresholds, and markings for exceeding or failing to meet the three core indicators: average drug cost per inpatient visit, antibiotic percentage, and essential drug usage percentage. The oncology, respiratory, and ICU departments are clearly marked as key departments for annual drug cost monitoring. For example, the page visually presents that the hospital's total drug cost in 2025 was 234.29 million yuan, with a year-on-year growth rate lower than the control limit. Outpatient drug cost indicators were better than the annual target, while average drug cost per inpatient visit and antibiotic percentage exceeded control standards, and the essential drug usage percentage did not meet the required standards.
[0087] The core indicator card display area uses independent data cards to display six key control indicators: total drug cost, outpatient drug cost, inpatient drug cost, average drug cost per outpatient visit, and average drug cost per inpatient visit. Each card indicates the annual target value and the actual value for the current year, along with the difference, and is accompanied by three status indicators: met, exceeded, and better than the target. For example, the inpatient drug cost card indicates a target of 99 million yuan, while the actual value was 102.39 million yuan, showing a difference of 3.4% exceeding the target; the average drug cost per outpatient visit target was 128 yuan, while the actual value was 124.5 yuan, showing a difference of 2.7% better than the target.
[0088] The progress tracking table for core indicators is uniformly structured with five columns: indicator name, target value, actual value, gap, and progress status. It lists eight core medical insurance cost control indicators, each with a corresponding numerical difference, and categorizes progress evaluations into several types: exceeding the target, better than the target, meeting the target, trending steadily, slightly exceeding the target, requiring enhanced monitoring, and not meeting the target. The indicators covered include total drug costs, outpatient drug costs, inpatient drug costs, the increase in average drug cost per outpatient visit, the increase in average drug cost per inpatient visit, average drug cost per outpatient visit, average drug cost per inpatient visit, the percentage of antimicrobial drug expenditure, and the percentage of essential drug usage expenditure. All values correspond one-to-one with the data on the page cards. For example, the target upper limit for the percentage of antimicrobial drug expenditure is 7.2%, and the actual value is 7.8%, a difference of 0.6%, indicating a need for enhanced monitoring; the target lower limit for the percentage of essential drug usage expenditure is 45%, and the actual value is 42.34%, a difference of -2.66%, indicating that the progress has not met the target.
[0089] The second section is the drug cost control monitoring area, titled "Average Drug Costs per Outpatient and Inpatient Visit and Their Growth Trends," featuring a five-year time-series line chart from 2021 to 2025. The chart uses three curves: a blue curve representing average drug costs per outpatient visit, a green curve representing average drug costs per inpatient visit, and a yellow curve representing the target average drug cost per inpatient visit. The horizontal axis represents the year, and the vertical axis represents the cost range, comprehensively recording the actual values of average drug costs per outpatient and inpatient visit over five years and their changes relative to the control target reference line, visually demonstrating the fluctuation patterns of drug costs over many years. For example, the line chart clearly shows that from 2021 to 2025, the average drug cost per outpatient visit gradually declined year by year, consistently falling below the annual control target value, while the average drug cost per inpatient visit consistently exceeded the set control benchmark.
[0090] This page belongs to the annual summary display module of the system's visual dashboard. It receives the hospital-wide annual analysis data pushed by the multi-level data distribution module and automatically generates a complete summary report according to the annual monitoring cycle. The page has a built-in function to highlight indicators that exceed the standard. It supports drilling down from the hospital-wide annual overview page to retrieve detailed data on departments and individual drugs. The generated report can be directly downloaded and pushed to the reporting application module, providing year-round drug cost control data support for early warning and clinical intervention, and implementing closed-loop management of early warning, diagnosis, intervention and feedback.
[0091] like Figure 3 As shown, Figure 3 This is a schematic diagram of a cardiology department drug cost management analysis report, specifically the cardiology department drug cost management analysis report for March 2026, with the statistical period limited to March 1, 2026 to March 31, 2026. The upper right corner of the page features "Preview Report" and "Download Report" buttons, supporting online viewing and file export. The bottom of the page indicates the basis for the report's generation: it is generated based on departmental drug usage data, indicator threshold data, and a drug knowledge base, and is intended solely for management analysis and decision-making reference.
[0092] The first section is the core discovery and alert area. This section categorizes and marks four types of drug management risks in the department, using different icons to distinguish the risk types, as follows: High-percentage drugs in total cost: Piperacillin-Tazobactam for Injection, accounting for 52.3% of the department's total drug cost, exceeding half. High-contribution drugs in terms of cost increase: Amiodarone Hydrochloride Injection, accounting for 63.5% of the department's total cost increase, the core source of the department's cost increase. Drugs with double high risk: Meeting both the criteria of high percentage of total cost and high cost increase, including both Piperacillin-Tazobactam for Injection and Amiodarone Hydrochloride Injection. Newly added drugs to watch out for in rapid growth warning: Propofol Emulsion Injection, accounting for only 2.0% of the overall drug cost, but accounting for 23.0% of the cost increase, indicating a rapid short-term growth rate requiring early clinical intervention. This section provides a preliminary classification and refinement of individual drug risks, allowing managers to quickly identify key drugs for departmental control without viewing the full list.
[0093] The second section is the department's annual drug cost control progress indicator, displaying the cardiology department's annual drug cost control target and its monthly progress: the annual drug cost control target is 15 million yuan, and the total drug cost in March was 1.285 million yuan, only 8.57% of the annual target. The standard progress should reach 25%, and the page indicates an abnormal progress, prompting users to pay close attention to the department's drug cost data. The section also displays year-on-year growth rates, showing a year-on-year increase of 8.2% and a month-on-month increase of 12.8%. For example, the progress figures provide a clear indication of the department's monthly progress being behind schedule, while the dual-dimensional growth data reflects the continuous increase in the department's drug usage, providing a double warning of management risks.
[0094] The third section consists of two tables ranking drugs, one on the left and one on the right. The left table lists the top 5 drugs by total drug cost, with four columns: serial number, drug name, total drug cost, and percentage of total drug cost. It lists the top five drugs by cost in the cardiology department for the month. Legend: bold red rows represent drugs accounting for more than 50% of total drug cost, and yellow background rows indicate drugs appearing on both lists. For example, piperacillin-tazobactam for injection ranks first, costing 320,500 yuan that month, accounting for 52.3% of the department's total drug cost, and is also marked in yellow, appearing on both lists simultaneously.
[0095] The table on the right lists the top 5 drugs by growth rate. It includes five columns: serial number, drug name, total drug cost, drug growth amount, and growth percentage. The legend explains that bold red rows represent drugs with a growth percentage exceeding 50%, and yellow background rows indicate drugs appearing on two lists simultaneously. For example, amiodarone hydrochloride injection ranked first in the growth rate list, with an increase of 38,100 yuan in costs that month, contributing 63.5% to the overall growth. The bold red text highlights the high risk of growth.
[0096] like Figure 4 As shown, Figure 4 This system provides a visual dashboard for analyzing drug cost structure. It receives multi-dimensional datasets from the hierarchical data distribution module and displays them in a structured manner across four dimensions: individual drugs, ATC drug classification, clinical departments, and drugs covered by medical insurance policies. It is equipped with various visualization formats, including lists, bar and line charts, pie charts, and detailed tables. All indicators are highlighted with red indicators, and users can drill down to retrieve detailed data, enabling managers to pinpoint the core sources of drug cost growth.
[0097] For the first row, the left-hand panel displays the top 5 drug costs by total amount. This panel lists the top five drugs by annual cost in a table format, including drug name, total cost, year-on-year increase, and expense increase. It fully displays the data for all five drugs, highlighting those with high year-on-year increases in red, clearly showing the consumption scale and rate of increase of high-priced, key monitored drugs. For example, human serum albumin had an annual cost of 2.102 million yuan, a year-on-year increase of 118%, with an annual expense increase of 205,000 yuan, making it the single drug with the highest consumption amount in the entire hospital.
[0098] The right-hand panel displays a pie chart of the ATC (Active Toxic Drug Control) drug cost structure and the top three indicators. The upper section features a circular pie chart showing the percentage of costs for each of the five major ATC drug categories: systemic anti-infectives, anti-tumor and immunomodulatory agents, digestive and metabolic drugs, blood and hematopoietic organ drugs, and others. The pie chart indicates the cost percentage for each category. The lower section provides a summary of the top three drug categories, including the annual total cost, hospital-wide percentage, and overall cost growth rate for each category, quantifying the cost control pressure for different pharmacological classifications. For example, systemic anti-infectives account for 31.95% of the hospital's total drug costs, corresponding to a total cost of 12.3 million yuan, with an overall growth rate of 11.5%, making it the drug category with the highest percentage in the hospital.
[0099] Regarding the second row, the left-hand panel presents a combination of bar charts and line graphs showing the top 5 departments in terms of total drug expenditure. The chart uses two display methods: green bars represent the total drug expenditure of each department in 2026, and the red line represents the year-on-year growth rate of drug expenditure. The horizontal axis sequentially displays the five high-consumption departments: oncology, general surgery, ICU, respiratory medicine, and orthopedics. The bottom of the chart includes textual annotations explaining that the statistical period is from January to May of the current year, and highlighting that oncology and general surgery saw the most significant increases in expenditure. For example, the oncology bar is the tallest among all departments, corresponding to a year-on-year growth rate of 20.6%, making it a key department for management with both high drug expenditure volume and high growth rate.
[0100] The right-hand panel displays a detailed table of the top 5 departments' drug expenses. The table includes five columns: department name, total drug expenses for the year, drug expenses for the same period last year, annual increase, and year-on-year growth rate. It provides a complete comparison of data for the top five high-consumption clinical departments, highlighting the growth rate in red and allowing for a horizontal comparison of the differences in drug expense increases across departments. For example, the ICU's drug expenses this year were 3.282 million yuan, an increase of 592,000 yuan compared to the same period last year, representing a year-on-year increase of 22.0%, ranking first in the hospital in terms of expense growth.
[0101] Regarding the third row, the left side features a pie chart showing the proportion of drug costs by policy type. The chart categorizes drugs into four types based on their medical insurance policy attributes: negotiated drugs, regular drugs, centralized procurement drugs, and auxiliary / key monitored drugs. The chart clearly indicates the percentage of each category's total hospital cost, visually representing the cost structure distribution of various medical insurance policy-related drugs. For example, negotiated drugs account for 35% of the hospital's total drug costs, making them the category with the highest proportion of policy-related drugs.
[0102] The right-hand panel displays a detailed table of core indicators for each policy type. The table is divided into four dimensions: policy type, total drug cost, annual increase, and year-on-year growth rate. It provides complete cost data for three categories of drugs: negotiated drugs, conventional drugs, and drugs procured through centralized procurement. The negative growth figures for drugs procured through centralized procurement are highlighted in red, clearly distinguishing the rising or falling trends in drug costs under different policies. For example, the annual cost of drugs procured through centralized procurement was 6.27 million yuan, a decrease of 855,000 yuan, representing an overall year-on-year decrease of 11.2%, demonstrating the effectiveness of the centralized procurement policy in controlling costs.
[0103] In one embodiment, this application also provides a medical expense data processing system based on multi-source data fusion, comprising: The medical multi-source data acquisition module is used to collect various heterogeneous original medical business data such as in-hospital HIS medication, electronic medical records, test results, and medical insurance policies; The control threshold benchmark configuration module is used to set adaptive early warning thresholds for costs, proportions, and growth rates based on the cost control requirements of DRG / DIP medical insurance payment reform and the clinical rational drug use evaluation standards. It also defines multiple irrational drug use quantitative judgment benchmarks, such as CMI-cost elasticity coefficient, dose deviation rate, treatment course exceedance value, and dispersion of actual and predicted costs. The medical data standardization preprocessing module is used to perform data cleaning, missing value completion, and feature engineering on multi-source heterogeneous medical data in sequence to build a standardized basic dataset of medical expenses. It extracts six feature dimensions: cost, drug, department, disease, time series, and medication guidelines to support the identification of cost anomalies and the diagnosis of medication behavior. The dual-time-series intelligent cost analysis module is used to perform cost time-series calculations by adopting an absolute threshold early warning + relative value attribution analysis mode and integrating ARIMA and LSTM dual-time-series prediction models. It breaks down the abnormal contribution from multiple dimensions such as department, drug category and disease to achieve accurate source tracing, and matches the data output update cycle of the visual dashboard and automatic report. The hierarchical data distribution closed-loop management module is used to split datasets and divide them into multi-level management units according to the dimensions of the whole hospital, departments, drug categories, and diseases. It encapsulates batches of analytical data containing cost indicators, medication abnormality tags, compliance judgment results, and intervention plans, and pushes them synchronously to the early warning, intervention, and reporting application modules to achieve a closed-loop management of early warning, diagnosis, intervention, and feedback.
[0104] The computer-readable storage medium provided in the above embodiments of this application and the medical expense data processing method based on multi-source data fusion provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.
[0105] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the medical expense data processing method, system, electronic device, and readable storage medium based on multi-source data fusion are basically similar to the embodiments of the medical expense data processing method based on multi-source data fusion described above, and are therefore described simply. Relevant parts can be referred to in the descriptions of the embodiments of the medical expense data processing method based on multi-source data fusion described above.
Claims
1. A medical expense data processing method based on multi-source data fusion, characterized in that, include: Collect drug cost-related data from the HIS system, including drug name, dosage, amount, and department affiliation information; Based on the cost control requirements of DRG / DIP medical insurance payment reform and the clinical rational drug use evaluation criteria, dynamic early warning thresholds for costs, proportions, and growth rates are set, as well as quantitative judgment thresholds for irrational drug use and the range of dispersion judgment between actual and predicted costs are determined. Preprocessing of multi-source heterogeneous medical data, data integrity verification and format standardization, construction of a standardized basic dataset of medical expenses, extraction of feature dimensions such as cost, drug, department, disease, time series, and medication guidelines, for use in cost anomaly identification and medication behavior diagnosis; Combining the needs of anomaly identification, medication diagnosis, hierarchical control, and full-process closed-loop management, an absolute threshold early warning + relative value attribution analysis mode is adopted. An intelligent cost analysis strategy is constructed by integrating ARIMA and LSTM dual time series prediction models. Anomaly contribution is traced through multi-dimensional decomposition, and the data output update cycle of the visual dashboard and automatic report is matched. Data sets are split into multiple levels of management units based on hospital-wide, departmental, drug category, and disease type. This generates batches of analytical data that include cost indicators, medication anomaly labels, compliance judgment results, and intervention plans. These data are then simultaneously pushed to the early warning, intervention, and reporting application modules to support closed-loop management of early warning, diagnosis, intervention, and feedback.
2. The method of claim 1, wherein, Based on the cost control requirements of DRG / DIP medical insurance payment reform and the clinical rational drug use evaluation criteria, dynamic early warning thresholds for costs, proportions, and growth rates are set, as well as quantitative judgment thresholds for irrational drug use. The range of dispersion between actual and predicted costs is determined, including: Combining the mandatory requirements for hospital cost control under the DRG / DIP medical insurance payment reform and the clinical rational drug use management and evaluation standards, based on the department's historical cost data, the ARIMA model is first used to generate baseline prediction intervals for costs, drug proportions, and cost growth rates. Then, the LSTM model is introduced to incorporate medical insurance policy identification data to complete residual correction. Combining annual cost control targets, CMI disease group weights, and medical insurance adjustment coefficients, multi-dimensional parameters are iterated to output multi-dimensional adaptively adjustable early warning thresholds. Based on the built-in quantitative evaluation indicators of CMI-cost elasticity coefficient, dose deviation rate, and treatment course exceedance value in the professional diagnostic knowledge base, standardized judgment rules for the behavior of drug use without indication, drug use exceeding the dose, and drug use exceeding the treatment course are constructed. A combination of ARIMA and LSTM time-series prediction models is used to perform time-series cost fitting operations, calculate the degree of dispersion between actual cost values and model prediction values, and set the critical boundary of dispersion as the basis for identifying cost anomalies. It is equipped with a combined identification mode of absolute threshold early warning and relative value attribution analysis, and combines the contribution of dispersion anomalies from multiple dimensions such as department, drug category and disease to support the accurate tracing of abnormal drug cost issues.
3. The method of claim 2, wherein, Preprocessing of multi-source heterogeneous medical data includes data integrity verification and format standardization. A standardized basic dataset of medical expenses is constructed, extracting features related to cost, medication, department, disease, time series, and medication guidelines for use in anomaly detection and medication behavior diagnosis. In response to the original data defects of multi-source heterogeneous medical data, such as HIS medication data, electronic medical records, test results, and medical insurance policies, including data disorder, missing indicators, and inconsistent business fields, we uniformly perform integrity verification and format standardization operations on various types of original medical business data to remove invalid data and unify the field formats of multiple systems. In conjunction with the requirements of DRG / DIP medical insurance cost control analysis, the target of drug cost anomaly identification, and the diagnostic scenario of irrational drug use, a standardized dataset construction specification is established, along with cross-business system data alignment rules and unified indicator conversion logic. Based on a multi-dimensional business feature extraction mechanism, a standardized medical expense dataset construction strategy was implemented, extracting six major feature dimensions: cost, drugs, departments, diseases, time series, and medication guidelines.
4. The method of claim 1, wherein, Combining the needs of anomaly identification, medication diagnosis, tiered control, and closed-loop management throughout the entire process, an absolute threshold early warning + relative value attribution analysis mode is adopted. An intelligent cost analysis strategy is constructed by integrating ARIMA and LSTM dual time-series prediction models. Multi-dimensional decomposition enables the tracing of anomaly contribution, and the data output update cycle is matched with a visual dashboard and automatic reporting, including: Combining the business needs of abnormal drug cost identification, intelligent diagnosis of irrational drug use, hierarchical management of the whole hospital and departments, and closed-loop management of early warning, diagnosis, intervention and feedback, the system calls the pre-processed standardized medical cost dataset based on the joint identification mode of absolute threshold early warning and relative value attribution analysis. It matches the three dimensions of department, drug category and disease to analyze and decompose the boundary, locate the cost statistics node corresponding to multi-source medical data, and divide the calculation range of cost anomaly contribution under each dimension. The system categorizes and labels drug cost indicators, clinical drug use behavior, medical insurance compliance information, and time-series consumption data across various dimensions. It integrates ARIMA and LSTM dual time-series prediction models to calculate the dispersion between actual cost values and model prediction values. It uses quantitative indicators such as CMI-cost elasticity coefficient, dose deviation rate, and treatment course exceedance to determine irrational drug use. It is compatible with visual dashboards and intelligent analysis reports with fixed update cycles, ensuring complete data correlation for indicators highlighted in red, multi-level data drill-down, and automatic report push. Establish a multi-dimensional anomaly tracing and visualization output linkage mechanism. After the dataset completes the multi-dimensional anomaly contribution decomposition calculation of departments, drug categories, and diseases, it generates hierarchical control analysis results and completes time-series arrangement according to the preset monitoring cycle. The results are then pushed to the front-end visualization application module to display the dynamics of drug cost control for the entire hospital, specialties, and single products in real time.
5. The method of claim 4, wherein, Data sets are broken down by hospital-wide, departmental, drug category, and disease type, and divided into multi-level control units. This generates batches of analytical data containing cost indicators, medication anomaly tags, compliance assessment results, and intervention plans. These batches are simultaneously pushed to the early warning, intervention, and reporting application modules, supporting closed-loop management of early warning, diagnosis, intervention, and feedback. Based on DRG / DIP medical insurance cost control, tiered clinical management and control and multi-module data carrying capacity, the top-level goal of dataset splitting is to support the data tiered distribution system that supports the closed-loop management of drug costs. The hospital is divided into four primary control units based on four dimensions: hospital as a whole, departments, drugs, and diseases. These units are further divided into secondary data encapsulation units for cost collection, anomaly marking, compliance accounting, and intervention plan matching. A progressive, hierarchical distribution architecture is established, consisting of an overall objective, a primary control unit, and a secondary encapsulation unit. This architecture clarifies data input at each level, hierarchical relationships, multi-module push notifications, and timing constraints. Complete batches of drug cost analysis are encapsulated and synchronously pushed to the early warning, intervention, and reporting modules. This architecture leverages multiple clinical intervention methods to achieve closed-loop management of early warning, diagnosis, intervention, and feedback.
6. A medical expense data processing system based on multi-source data fusion, characterized in that, The system is used to execute executable instructions to perform the medical expense data processing method based on multi-source data fusion as described in any one of claims 1 to 5.
7. An electronic device, comprising: include: First processor; The processor also includes a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the medical expense data processing method based on multi-source data fusion as described in any one of claims 1 to 5 by executing the executable instructions.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the second processor, it implements the medical expense data processing method based on multi-source data fusion as described in any one of claims 1 to 5.
